Particle.news
Download on the App Store

AI Pair Programming Dulls Code Review and Learning, Study Finds

The observational study of 19 programmers will be presented at ASE 2025 in Seoul.

Overview

  • Developers working with GitHub Copilot were more likely to accept suggestions without critical evaluation, the researchers report.
  • Human–AI sessions produced fewer and narrower knowledge-transfer interactions than human–human pairs, often staying confined to immediate code.
  • The authors warn that uncritical trust in assistants can foster complacency and contribute to technical debt that surfaces later.
  • AI helpers proved useful for routine tasks and reminders, yet fell short of the richer collaboration needed for complex problem solving.
  • The work by Saarland University’s software engineering group, funded by the ERC ‘Brains On Code’ grant, will be presented by first author Alisa Welter.